Tuesday, April 24, 2018

Frankencomputer for deep learning: running ubuntu 16.04 on a Dell Precision T5500

Choosing a second hand computer:

As my previous laptop died of heat, I was looking for a new machine with a tight budget. Having read the post from Denys Katerenchuck on a budget friendly PC for deep learning it was tempting find an old machine capable of powering a gpu such a GTX 960 or better.
Old workstations are capable of powering such GPU. The Dell Precision T series are valuable since they comes with a 875 W PSU in the case of the T5500.

A lot of peoples tweak this model  for gaming (thanks for the link), for example this one. Several videos show such Dell Precision workstations running games:
    I bought on Ebay a Dell Precision T5500 (12 Gb RAM, xeon 5520 CPU, no HD, no cable, 875W PSU) for 133€ (~100€ for the machine and ~30€ for the transport from UK to France).
    3x4 Gb RAM

    PCI slots

    When first connected to a TV as monitor (no HD at this stage), an error message was displayed (due to the size of the screen?):

    Fortunately, initial tests show no hardware problems:

    A 128 Gb SSD from a previous laptop computer was connected on SATA1:

    Ubuntu 16.04 was installed from a usb-key. Installing Ubuntu took 15 minutes.
    Some BIOS settings were modified to boot on the SSD. SATA-1 was checked and other settings were unchecked:

     RAID config was switched from "RAID-on" to "RAID auto detect/ AHCI":
    Powering up looks like:

    If needed, the boot-info is here.


    This Precision T5500 model comes with a 0D883F motherboard accepting E55xx or E56xx CPU, this is not the best possible choice for a T5500, currently, it has one Xeon E5520 CPU.

    GPU card:

    The graphic card provided with the computer can't be used for deep learning, even not for 3D with geogebra !! It's a:
     VGA compatible controller: Advanced Micro Devices, Inc. [AMD/ATI] RV635 [Radeon HD 3650/3750/4570/4580]

    in 2016, Denys Katerenchuck recommended the GTX 960 4Gb as a good tradeoff between price and performances. In 2018 such card can be found on ebay for less than 150€. The Dell T5500 can power a GTX 1050 from its PCI slot. It has also a six pins power connector (left) to power an EVGA GTX 1060 for example.

    There's  an interresting post regarding powering GPU and 6/8 pins cable. Again, according to Akshat Verma (see the answer about T5500) it seems possible to plug a GTX 1070 in a T5500.
    Somes reported that they installed a GTX 960 and even a 1080.


    Dell splash screen visible on first boot
    The bios version is a A02 from 2009.  Dell provides updates for A02 both for windows and linux (for redhat, not ubuntu). It seems that the BIOS has to be upgraded in order to install a recent GPU.
    Questions about BIOS version and GTX960 can be found, The BIOS upgrade is a recurrent question.

    BIOS version history


    - no title specified

    Fixes & Enhancements

    07 Jun 2013
    1.Enhanced Broadcom onboard NIC 5761 support.
    2.Enhanced CCTK support for Turbo Boost setting.
    27 Feb 2013
    1.Enhanced PSA support.
    2.Update PSA to A4527.
    3.Enhanced Dual CPU memory configurations support.
    04 Dec 2012
    1.Enhanced Windows 8 support.
    07 Jul 2012
    1.Add High IO performance option in Setup.
    2.Disable QPI L1 when C-States have been disabled.
    3.Add the changes that ties ASPM to the C-State switch.
    4.Add TSEG protection feature ENABLE_TSEG_SECURITY.
    07 Jul 2012
    1.Update new RAID OROM v10.8.0.1303.
    03 Nov 2011
    1.Enhanced TPM support.
    03 Nov 2011
    The following changes have been made to BIOS A09 to A10:

    1. Added support for newer Processors.
    2. Added support for bridged AGP video cards.
    3. Updated the copyright to 2003.
    4. Fixed UDMA support for 48-bit LBA hard drives (> 137 GB).
    5. Fixed problem where system would occasionally shut off when Ctrl-Alt-Del is pressed.
    6. Free up unused portions of E000:0000 to E000:FFFF memory for memory managers to use.
    07 Jul 2012
    1.Added new Hard Disk master password algorithm support.
    2.Added support to install single graphic card in 2nd PCIEx16 slot.
    3.Enhanced PCIE Slot interrupt handling.
    4.Enhanced the memory map algorithm for 128GB configuration.
    07 Jul 2012
    1.Enhanced method to check the failure memory
    2.Updated ACPI SRAT table information
    07 Jul 2012
    1. Enhanced security device support
    07 Jul 2012
    1. Updated Intel Xeon?Processor 5600/3600 Series microcode to rev 10.
    2. Updated Intel Xeon Processor 5500/5600 Platform Reference Code to revision P2.91.
    3. Enhanced TPM remotely provision.
    4. Enhance VT-d
    07 Jul 2012
    1. Updated to the latest Intel (R) Xeon (R) Processor 5600 Series microcode
    07 Jul 2012
    1. Supported E5620, E5630, E5640, X5650, X5660, X5667, X5670, X5677, X5680, W3680 CPUs.
    2. Updated Intel Xeon?500/5600 Platform Reference Code to revision P2.7
    3. Removed S1 support.
    4. Removed "Optional HDD fan" support.
    5. Reported riser's DIMMs information(asset tag, serail number) in SMBIOS.
    6. Used the same fan setting before and after S3.
    7. Enhanced the compatibility with certain PCI-Express Gen1 cards.
    8. Enhanced NUMA under RHEL5.3.
    9. Removed HDD Acoustic support.
    07 Jul 2012
    1. Added error detection for bad monitors or cables when entering setup menu.
    2. Enhanced PCIe bar allocation.
    3. Enhanced algorithm for Graphics card with multiple OPROMs.
    4. Updated BIOS fan descriptions.
    5. Fixed BIOS cannot boot from SATA CD/DVD when setting USB controller to "No Boot" in BIOS setup menu.
    07 Jul 2012
    1. Implemented fix for intermittant boot issue with 6.4 GT/s CPUs.
    2. Added feature to display a message if the DIMM configuration is not optimal.
    3. Updated Intel(R) Memory Reference Code.
    4. Added microcode update revision 11 for Nehalem D0-step.
    5. Fixed boot issue with RAID and ATI FireMV 2450.
    6. Added TCM support.
    7. Added updated IO programming.
    8. Fixed possible hang condition when VT-d is enabled.
    9. Moved "Optional HDD Fan" to the "Post Behavior" section in Setup.
    10. Updated SMBIOS tables.
    11. Added support for Windows 7.
    12. Updated fan settings.
    13. Corrected memory channel information in Setup.
    14. Improved the allocation of system resources.
    Testimony of a user showing he had a hard time when he tried to upgrade the firmware. The user reported that : "Oh I should state that when I got the machine it had bios version A5 on it. I was able to get to A9, but I had to flash each BIOS in order to get to that point."
    Different feedbacks from the web, advice to upgrade the BIOS directly from the current version to the last version

    Upgrading from windows yields:

    From Ubuntu or freeDos:

    There are different possible ways to upgrade the BIOS:
    By the way, The T5500 refused to boot from a freedos usb stick up to now ...

    Motherboard versions 

    It seems  that the motherboard can comes in different flavours:
    • D883F
    • CRH6C
    • W1G7K 
    The most capable motherboard seems to be the CRH6C model supporting up to X5690 xeon CPU. One user reported the use of two X5660 on the D883F model with the latest BIOS (shoud be A16).

     Upgrading the GPU

    An asus GTX 960 4Gb turbo was waited:

    The card is long for a T5500, some place has to be made to plug it in the case:
    The blue plastic locker was removed
    The hard disc support was partially cut and bent.
    As the T5500 was powered on, it just boot (from a BIOS A02!) and open a linux session (with a nouveau driver):

    So no firmware update!

    Some softwares were installed (nvidia ppa, driver 384-130...) and this yields:
    the sensors yields the same info regarding the GPU temperature:

    CUDA 9 + cuDNN 7 on ubuntu 16.04

    Several tuto are available to install cuda 9 + cudnn:
    After installation, the nvidia driver was updated to 396.26:

    Checking install

    Copy the cuda sample files somewhere to  the home directory and build them:

    Launch deviceQuerry for example:

    The GTX 960 card (PCIe 3) is plugged on a PCIe 2 16x slot. How the card and the computer communicate? May be bandWidthTest can bring some insight:

    Sunday, February 18, 2018

    Training OverlapSegmentationNet with GPU acceleration on google-colaboratory

     OverlapSegmentationNet is a python implementation of Unet intented to resolve pairs of overlapping chromosomes by semantic segmentation (See previous post).

    OverlapSegmentationNet can now be trained on the google colaboratory platform with GPU acceleration.

    The whole files and data are available from google drive. Wouldn't have been able to set-up without the following article.

    Thursday, October 12, 2017

    Predictions from OverlapSegmentationNet after 14 epochs

    OverlapSegmentationNet is a UNet model implemented with Keras. The network was trained for 14 epochs (~8 hours on a GT740M GPU), this is not enough to make a good prediction but it is enough to play with the model to understand how a prediction from a single image looks like.


    The dataset used for training was the first proposed on kaggle. It consists in 13434 pairs of greyscale/ground truth images of size 88x88. Each greyscaled image has a pair of overlapping chromosomes. The overlaps were synthetically generated from images of single chromosomes:

    Making a prediction from a single greyscaled image:

    The aim is to predict the segmentation labels with a Keras model (OverlapSegmentationNet). The 88x88 greyscaled image must be converted into a 4D numpy array (1, 88, 88, 1).
    A prediction is then 4D array of shape ( 1,88,88,4), since there's one component for each label (background, chromosome 1, chromosome2, overlap). A prediction looks like:
    where each component is a floating-point image. With another chromosome, we have:
    A prediction takes ~ 40 ms for one low resolution image.

    Prediction made on an image which do not belong to the train dataset:

    A prediction was made on an image belonging to a validation dataset:
    To be used for prediction, the image was down sampled and recut to fit 88x88 (and its corresponding ground truth segmentation too):
    Clearly, with so few epochs the prediction is bad:
    The one-hot encoded prediction can be thresholded and back converted into a multilabel segmentation and compared to the ground truth segmentation:

    Notebook is here:

    More details are available in the notebook:


    Friday, August 25, 2017

    Training OverlapSegmentationNet with gpu acceleration on a Ubuntu powered laptop.


    Harnessing the gpu 1 : nvidia driver

    It is  supposed that a fresh ubuntu 16.04 is set-up.
    Install Nvidia proprietary driver from the system parametres menu:
    The NVidia installed files, with synaptic package manager, are:

    Nvidia acceleration can be switched on graphically after having installed nvidia-settings :

     Check driver installation from a terminal by starting glxgears:

    When NVIDIA mode is selected, glxgears performances should increase.

    Harnessing the gpu 2: install CUDA 8, cudnn 5.1

    CUDA8 was installed from a .deb file available from nvidia's site:

    Check cuda installation using the "./deviceQuerry" method (see the end of the post).

    The Cudnn 5.1 library was installed by copying the files at the right places (Installing cudnn 6 from .deb archive wasn't a good idea. )


    Install Tensorflow and Keras 

    There's different ways to install tensorflow. Here, both tensorflow with gpu support and keras were installed  in a virtual environnement with pip.

    Tensorflow 1.2.1 and Keras 2.0.6 were installed .

    h5py was installed from pip too.
    Don't forget at the end from a terminal to :

    $source .bashrc

    Installation of OverlapSegmentationNet:

    OverlapSegmentationNet is an implementation of UNet by Dr Hu .

    The image-segmentation-chromosomes directory contains:
    $ ls
    code  images    README.md


    The first version of the low resolution dataset was used. Download it and move it into the code directory:
    $ ls
    create_histogram_visualisation.ipynb  processInputImages.py~
    Explore-data.ipynb                    __pycache__
    Explore-model2.ipynb                  sub-trainModel.py
    histogram.png                         trainModel.py
    LowRes_13434_overlapping_pairs.h5     trainModel.py~
    models                                utilities.py
    OverlapSegmentationNet.py             utilities.py~
    OverlapSegmentationNet.pyc            utilities.pyc
    preprocessing-jp.ipynb                xdata_88x88.npy
    processInputImages.py                 ydata_88x88_0123_onehot.npy

    In a terminal,  run processInputImages.py as follow:
    $python processInputImages.py
    The code will generate the two files highlighted in green above.


    Switch into gpu mode with (then logout / login):

    Activate the proper virtual environnement. Mine was called tfgpu, so for my computer, it is:
    $source VirtualEnv/tfgpu/bin/activate
    Edit the code to set-up the desired number of epoch
    From a terminal, run trainModel.py. For example in my computer, it is:

    (tfgpu) jeanpat@WA50SHQ:~/image_segmentation_chromosomes/code$ python trainModel.py
    If the gpu is properly used by tensorflow the terminal should yield something like:

    The usage of the gpu is mentionned, furthermore the cpu usage is not stucked on 100%.

    On a modest nvidia GT740M gpu, 14 epochs were run for 8~9 hours.

    Saturday, July 1, 2017

    Counting impacts on a target (Traces of heavy ions beam) with the help of kmeans clustering.

    When studying effect of ionizing radiation (heavy ions with high LET) on cells, the fluence (particules/cm²) of the particles beam must be known. Detectors in the GANIL (Caen, France) particles accelerator can estimate the fluence. However, when irradiating cells, radiobiologists (CEA CIRIL) need to localize the impacts of the ions and to count the number of impact per cell.

    The application TRACES (scripts for the aphelion application) was developped: It was capable of recording and saving images from a camera, segmenting and counting traces: 
    The application was designed to analyse images as the following:

    raw image of traces

    Counting impacts  in a python notebook:

    Contrary to the TRACES application, the following jupyter notebook can't capture images (It should be possible to do it in a python script with the cv2 module) but it can perform some automatic classification.
    Distinguishing traces resulting from one or several impacts is a classification issue which can be explored with unsupervised classifier such k-means clustering: